Chemosphere 292 (2022) 133422

Contents lists available at ScienceDirect

Chemosphere
journal homepage: www.elsevier.com/locate/chemosphere

Computational methods on food contact chemicals: Big data and in silico

screening on nuclear receptors family
Pietro Cozzini a, *, Francesca Cavaliere a, Giulia Spaggiari a, Gianluca Morelli b, Marco Riani b
a
Molecular Modelling Lab, Department of Food and Drug, University of Parma, Parco Area Delle Scienze 17/A, 43124, Parma, Italy
b
Department of Economics and Management and Interdepartmental Center of Robust Statistics, University of Parma, Via J. F. Kennedy 6, 43100, Parma, Italy

H I G H L I G H T S G R A P H I C A L A B S T R A C T

• Molecular docking and robust consensus

scoring are useful to identify possible
food and water dangerous molecules.
• Endocrine disruptor prediction using in
silico methods to save time and cost.
• Database and big data approaches to
accelerate hazard identification.

A R T I C L E I N F O A B S T R A C T

Handling Editor: A. Gies According to Eurostat, the EU production of chemicals hazardous to health reached 211 million tonnes in 2019.
Thus, the possibility that some of these chemical compounds interact negatively with the human endocrine
Keywords: system has received, especially in the last decade, considerable attention from the scientific community. It is
Computational chemistry obvious that given the large number of chemical compounds it is impossible to use in vitro/in vivo tests for
Consensus prediction
identifying all the possible toxic interactions of these chemicals and their metabolites. In addition, the poor
Database
availability of highly curated databases from which to retrieve and download the chemical, structure, and
Nuclear receptors
Toxicology regulative information about all food contact chemicals has delayed the application of in silico methods. To
overcome these problems, in this study we use robust computational approaches, based on a combination of
highly curated databases and molecular docking, in order to screen all food contact chemicals against the nuclear
receptor family in a cost and time-effective manner.

1. Introduction
A research project starts with a question. The main question of this
project is: how we can evaluate all the possible food contact chemicals
against a protein family to discover potential endocrine disrupting
activity. It is obvious that, given the large number of chemical com­
pounds and their metabolites existing and developed every year, it is
impossible to use in vitro (or in vivo) tests for identifying all possible toxic
interactions. The solution is to use computational approaches to reduce
the number of wet tests, seeking only the most probable interactors.

* Corresponding author.
E-mail addresses: pietro.cozzini@unipr.it (P. Cozzini), francesca.cavaliere@unipr.it (F. Cavaliere), giulia.spaggiari@unipr.it (G. Spaggiari), gianluca.morelli@
unipr.it (G. Morelli), marco.riani@unipr.it (M. Riani).

https://doi.org/10.1016/j.chemosphere.2021.133422
Received 25 October 2021; Received in revised form 20 December 2021; Accepted 22 December 2021
Available online 28 December 2021
0045-6535/© 2021 Elsevier Ltd. All rights reserved.
P. Cozzini et al.
Endocrine disrupting chemicals (EDCs) are exogenous substances
that can interfere with the synthesis, secretion, transport, binding, and
elimination of natural hormones in the body that are responsible for the
maintenance of homeostasis, reproduction, and behavior (Kavlock et al.,
1996). Human exposure to EDCs occurs through oral consumption of
food and water, contact with skin, inhalation, or intravenous, route
(Kabir et al., 2015). These molecules are highly heterogenous and
include pesticides, plasticizers (i.e., phthalates, bisphenols), persistent
organic pollutants (POPs) (i.e., dioxins, polychlorinated biphenyls), but
also chemicals added to food to enhance some characteristics (i.e., fla­
vourings, food additives), or naturally occurred, such as mycotoxins.
EDCs can act through different mechanisms: mimicking the action of a
naturally produced hormone, blocking hormone receptors in cells,
interacting indirectly by influencing the biosynthesis or availability of
normal hormones. Between them, the most privileged route is the
interaction with nuclear receptors (NRs). Nuclear receptors are a su­
perfamily of 48 ligand-activated transcription factors, including estro­
gen receptor (ER), androgen receptor (AR), mineralocorticoid receptor
(MR), glucocorticoid receptor (GR), progesterone receptor (PR), and
thyroid receptor (TR). NRs share a common structural organization
composed of an N-terminal region (A/B domain), a conserved region
DNA-binding domain (DBD), and a ligand-binding domain (LBD)
responsible for ligand recognition. The alteration of nuclear receptors
pathways is correlated to many pathologies, such as breast cancer,
prostate cancer, and testicular cancer, infertility, cardiovascular com­
plications, disturbances in energy metabolism, immune responses,
impairment of cognitive functions and the regulation of cell prolifera­
tion and differentiation, hypertension, obesity, and so on (Dall’Asta,
2016) (De Coster and Van Larebeke, 2012) (Desvergne et al., 2009)
(Fucic et al., 2012) (Luccio-Camelo and Prins, 2011) (Odermatt and
Gumy, 2008) (Petrakis et al., 2017) (Safe, 2004) (Schug et al., 2011)
(Gore et al., 2015). In order to prevent human diseases, in the past de­
cades, different regulatory and policy approaches were made even if the
identification and safety assessment of potential EDCs is complicated
both by the observed low-dose effects and the often long-term exposure
or exposure during a critical window early in development. One of these
is the REACH (Registration, Evaluation, Authorisation, and Restriction
of Chemicals) legislation that is committed to protecting human health
and the environment from hazardous chemicals. However, testing all the
possible EDCs against all the potential targets is very important but also
an expensive, long and difficult task (e.g., the nuclear receptors family
contains 48 members). In fact, these tests are still mainly based on
biological and animal experimentations (toxicity tests), very time- and
cost-intensive, and which cause millions of animals’ death every year. In
this context, in silico methods, already well-established tools in drug
discovery, can be good tools either in the identification of new EDCs or
pointing in the right direction when finding the mechanism of action for
already known EDCs. Computational approaches produce predictive
models that are more rapid and less costly than in vitro and in vivo tests,
allowing a large amount of data concerning numerous chemical sub­
stances to be generated and analysed in a short time without the use of
test animals (F. Cavaliere et al., 2020). A key prerequisite for the suc­
cessful application of computational modeling techniques is the quality
of the input data. The availability of open access databases offers the
capability to retrieve a huge amount of information from different data
sources. The CAS Registry Number (RN) has been chosen, long time ago,
as a unique and unambiguous numeric identifier for a specific chemical
compound. It is developed by the American Chemical Society to help
scientists to retrieve and use information from different data sources.
Since it may be unique, validated, and internationally recognized, the
governmental agencies rely on CAS RNs for substance identification.
However, CAS RNs are often used improperly by the scientific commu­
nity and there is no check made by the American Chemical Society.
Thus, it is really common to find some errors and this wrong information
propagates easily across the Internet (Grulke et al., 2019). In fact, con­
flicts in the chemical identifier are not so rare in public resources and
2
Chemosphere 292 (2022) 133422
these errors propagate quickly and easily across the internet. These
undermine the effort of in silico methods. So far, much attention has been
paid to structure normalization to ensure the detection and the correc­
tion of three-dimensional errors and a variety of public and commercial
toolkits exist to address this problem. However, less attention is often
given to the consistency of the association between chemical identifiers
(CAS RN and name) and chemical structures. For example, the com­
pound classified as flavouring having the CAS N: 563187-91-7 and the
common name “l-Menthone-1,2-glycerol ketal” in the EFSA list is a
typical example of CAS:Name wrong association. In fact, this CAS
actually corresponds to “DNA (mouse strain C57BL/6 J clone
5430425J12 EST (expressed sequence tag))” and the correct CAS RN of
the compound “l-Menthone-1,2-glycerol ketal” is 67785–70-0. More­
over, although CAS RN is commonly used as an identifier of the majority
of databases, in several databases molecules are classified using different
identifiers and thus there is often a lack of standardisation (Hersey et al.,
2015). Although data quality is undoubtedly important for every data­
base, they may have been developed with different aims and scope, and
it is unreasonable to expect the same degree of curation. The increasing
amounts of compounds released every year (500–1000 new molecules)
and that are in contact with food, along with the different sources of
data, have made it difficult to check manually the reliability of data. In
view of this, it is essential to design and implement a data curation
pipeline into an automated procedure.
A wide number of computational applications (tools) specifically for
the analysis of EDCs are available in the literature in order to determine
the relationship between one compound and its toxic effect. In partic­
ular, the molecular docking technique is a well-establish application to
study protein-ligand interaction, which means analysing if the ligand
has the suitable physical-chemical characteristic, shape, the volume to
fit properly into the binding cavity of the receptor. Molecular docking is
mainly composed of two main parts: an algorithm that is used to predict
different binding poses of a molecule in the protein binding site, and a
scoring function used to evaluate the strength of ligand-protein inter­
action, i.e., to predict its binding affinity. Different algorithms and
scoring functions exist but answering the question of which algorithm or
scoring function is the best one, is a complicated task (Morris and
Lim-Wilby, 2008). In fact, each docking software (that is the sum of
algorithm and scoring function) has been trained with different proteins
and ligands. Thus, before starting a molecular docking analysis, it should
be advisable to identify the more appropriate software based on the
trained protein-ligand complexes that best fit with the proteins and li­
gands under investigation. However, in the present work, 31 different
nuclear receptors with different binding pocket characteristics and a
huge number of heterogeneous molecules from a chemical and struc­
tural point of view were considered. Thus, it is unthinkable to identify a
single docking program that may have the same performance for all
nuclear receptors and for all food contact molecules. For that reason, we
used a robust consensus scoring approach using two different docking
software and four different scoring functions. The combination of more
scoring functions allows to reduce the number of false-positive and to
obtain more reliable results by compensating the deficiencies of each
scoring function, leading to an improvement of the performances (Ter­
amoto and Fukunishi, 2007) (Wang et al., 2003). Such as Bissantz and
co-workers have highlighted, the use of three different scoring functions
enhances the capability to reach hit rates from 10% up to 70% (Bissantz
et al., 2000).
The goal of this work is to predict a possible endocrine disrupting
activity of a huge set of molecules that can contact the food as a base for
further in vitro/in vivo tests using computational methods that do not
consider the intake dose. The following approach takes into consider­
ation the interaction between a ligand (i.e. the endocrine disruptor
compound) and the binding site of a receptor (i.e. the nuclear receptor)
that is considered the molecular initiate event (MIE). This event is
fundamental from a biological point of view because it is the first
mechanism that, in most cases, initiates a biological effect based on the
P. Cozzini et al. Chemosphere 292 (2022) 133422

occurrence of conformational changes, signaling cascade as well as Table 1

interaction with other proteins. However, molecular docking does not SQL and NoSQL database structure definition.
predict the binding affinity of a ligand to a protein unless a correlation Field Data type (SQL) Mapping (NoSQL)
analysis was made using known experimental binding affinity (Kd, Ki,
CAS CHAR(16) keyword
Ka, etc.) and the docking score value. In fact, molecular docking and CID CHAR(20) keyword
scoring function refers to the binding interaction of a ligand for a protein EC number CHAR(50) keyword
that means analysing if the ligand has the suitable physical-chemical Common name TEXT text keyword
characteristic, shape, the volume to fit properly into the binding cav­ IUPAC name LONGTEXT text keyword
Molecular Formula CHAR(100) text
ity of the receptor. Thus, it predicts how strong is the interaction. Canonical SMILES TEXT keyword
Although it can be wrongly thought that if a ligand interacts more tightly InChI LONGTEXT keyword
with a receptor, it should have a high binding affinity, this concept is not InChIKey CHAR(254) keyword
so obvious since other mechanisms are involved in determining the MW FLOAT double
Volume FLOAT double
binding affinity. Lower binding force doesn’t mean low in take dose, it
logP FLOAT double
means a lower DG◦ of binding. Acceptor INT(3) byte
Donor INT(3) byte
2. Material and methods Chiral INT(3) byte
Hydrophobe INT(3) byte
Atom Count INT(3) short
2.1. Database resources Bond Count INT(3) byte
Ring Count INT(3) byte
Different databases and web sources have been used to identify the Positive Charge INT(3) byte
molecules that come into contact with food: European Food Safety Au­ Negative Charge INT(3) byte
Total Charge INT(3) byte
thority (EFSA) (www.efsa.europa.eu), United Stated Environmental
EFSA link CHAR(254) keyword
Protection Agency (EPA) (www.epa.gov), Food Packaging Forum (http ECHA link CHAR(254) keyword
://www.foodpackagingforum.org), and European Chemicals Agency .mol2 LONGTEXT keyword
(ECHA) (www.echa.europa.eu). Classification CHAR(100) text

2.2. Data quality

three different subgroups:
The entire procedure described below has been implemented as two
a) Chemical names: CAS, CID, EC number, common name, IUPAC
different Python procedures, with a common part used to check CAS RN
name;
validity. In fact, most public databases use Chemical names and CAS RNs
b) 1D chemical information: molecular formula, canonical SMILES,
as substance identifiers. CAS RN is widely used across scientific litera­
InChI, InChIKey;
ture, Internet resources, and the chemical regulatory domain. Data are
c) Chemical information: molecular weight, volume, logP value, num­
often stored using CAS RN as the primary key to the database and
ber of acceptor atoms, number of donor atoms, number of chiral
chemical names and synonyms as secondary identifiers. A CAS RN can
atoms, number of hydrophobic atoms, atom count, bond count, ring
be considered valid if it fulfils two rules: 1) it is composed by 3-numeric
count, rotational bond count, positive charge atoms, negative charge
parts separated by hyphens (## … - ## - #); 2) it satisfies the
atoms, total charge;
“checkdig” validation formula developed by CAS (www.cas.org/s
d) Regulative information: EFSA and ECHA link;
upport/documentation/chemical-substances/checkdig). CAS numbers
e) Three-dimensional structure in .mol2 format;
are preliminarily checked for the presence of leading zeros and zeros are
f) Classification: it classifies the molecule based on its use in the food
removed. After that, the checkdig formula has been used on CAS
industry: flavouring, pesticide, dioxin, etc.
numbers to verify their correctness.
The detail information about where and how these data have been
2.2.1. First procedure
obtained is explained below.
Using CAS RN as input query, the entire procedure can retrieve and
check data congruence of the InChIKey extracted from three different
2.3.1. PubChem information
servers: PubChem (www.pubchem.ncbi.nlm.nih.gov), ChemIDPlus
PubChem database has been used to retrieve some food contact
(www.chem.nlm.nih.gov/chemidplus) from the National Institute of
chemical data, as explained in the previous procedures: "CID", "Com­
Health (NIH), and CompTox Chemistry Dashboard (www.comptox.epa.
mon_name", "IUPAC_Name", "MolecularFormula", "MolecularWeight",
gov/dashboard) from EPA. Since the manual curation part of incon­
"CanonicalSMILES", "InChI", "InChIKey".
gruent data and/or unfound CAS RN took a great amount of time, a
second procedure has been developed.
2.3.2. ECHA number and ECHA link
A python script has been developed to convert CAS RNs into fixed
2.2.2. Second procedure
URLs to automatically retrieve EC numbers and to provide the corre­
Starting from the CAS RN information, it has been converted into
sponding link to the ECHA website’ Substance Infocard.
fixed URLs to automatically extract the correct InChIKey information
within the CAS database (www.commonchemistry.cas.org), which is the
2.3.3. 3D structures
official repository of CAS RN. In this step, the presence of salt and
The three-dimensional structures (in .sdf format) of molecules that
mixture was also checked. At the end, the InChIKey information was
passed the previous steps have been retrieved from PubChem using a
used as input query for extracting other information from PubChem
third python script.
("CAS", "CID", "Common_name", "IUPAC_Name", "MolecularFormula",
"MolecularWeight", "CanonicalSMILES", "InChI", "InChIKey").
2.3.4. Calculated chemical information
To store additional chemical information, other data have been
2.3. Database descriptors
calculated using two software:

The foodchem DB stores 27 different fields that can be divided in

3
P. Cozzini et al.
• Sybyl v.7.: Acceptor, Donore, Hydrophobe, AtomCOunt, BondCount,
RingCount, RotBonds, Chiral, logP value, Volume (Å3);
• FLAP: number of Charge – and Charge + and the Total Charge.
Moreover, the FLAP (Fingerprint for Ligand and Protein) software
was also used to convert the .sdf file into a .mol2 file.
2.4. SQL and NoSQL
The data have been organized into two different databases, MariaDB
and Elasticsearch, written implementing SQL and Bigdata technology
(NoSQL – Not only SQL) respectively. We decided to implement two
versions of the same database to answer two requirements. An SQL DB
storing structural data of the selected molecules, more suitable for
docking and molecular dynamics analysis, and a Big Data version able to
store a different kind of information, not only structural information but
also in vitro/in vivo tests, regulatory reports, etc. The specification of the
structure/mapping used in the present work is explained in more detail
in Table 1.
2.5. Protein preparation
The crystallographic structures of 31 nuclear receptors of Homo sa­
piens were downloaded from the Protein Data Bank (PDB) (www.rcsb.
org). Among them, only 26 structures with high reliability and quality
are available. For this reason, the nuclear receptors (3) with fragmented
portions, such as constitutive androstane receptor (CAR), nuclear
receptor-related 1 protein (NURR1), and estrogen-related receptor alpha
(ERRα), were built and minimized for 1 ns with NAMD 2.13 software
package. In addition, the mutated amino acids present in glucocorticoid
receptor (GR) (F602S) and steroidogenic factor 1 (SF-1) (C247S and
C412S) crystallographic structures were replaced. The receptor struc­
tures were processed using Sybyl software v8.1 (www.tripos.com).
Water molecules and ligands were removed, and hydrogen atoms were
added. Energy was minimized using the Powell algorithm with a
coverage gradient of ≤0.5 kcal (mol Å)− 1 and a maximum of 1500 cy­
cles. For the molecular docking with AutoDock (see below), the receptors
were further processed: using AutoDockTools software polar hydrogens
are added to the proteins and the Gasteiger charges were calculated to
assign AD4 type to each atom.
2.6. Ligand preparation
Structural coordinates of the endogenous and putative ligands were
retrieved from the NCBI PubChem compound database. Software FLAP
was used to assign the correct protonation state to each ligand (pH =
7.4).
2.7. Molecular docking with GOLD software
The GOLD software v5.8.1 (CCDC; Cambridge, UK; www.ccd.cam.ac.
uk) was applied in order to dock ligands into the binding site of the 31
nuclear receptors. For each compound and receptor, 30 binding poses
were generated. The binding site centroid of each receptor was defined
using the coordinates of the crystallographic complexes. The side chain
flexibility was allowed for each receptor amino acid. For the genetic
algorithm run, a maximum number of 100000 operations were per­
formed on a population of 100 individuals with a selection pressure of
1.1. The number of islands and the niche size were set to 5 and 2,
respectively. The default GoldScore fitness function was applied for
performing the energetic evaluations. The distance for hydrogen
bonding and the cut-off value for the van der Waals calculation were set
to 2.5 Å and 4.0 Å, respectively. Flip pyramidal N, flip amide bonds, and
flip ring corners were allowed for ligand flexibility options. After that,
all the poses generated by GOLD software were rescored using the
scoring functions ChemScore and HintScore (HINT, Hydropathic
4
Chemosphere 292 (2022) 133422
Table 2
The total number of food contact chemicals falling in each sub­
class. Food contact chemicals are divided into 11 subclasses:
dioxins, acrylamide, flavourings, food additives, furans, myco­
toxins, pesticides, phthalates, bisphenols, polychlorinated bi­
phenyls (PCBs), and food contact chemicals contained in the
database of Food Packaging Forum (FCCDB).
Classification Total number (8091)
Dioxins 75
Acrylamide 1
Flavourings 2091
Food Additives 110
Furans 133
Mycotoxins 327
Pesticides 465
Phthalates 361
Bisphenols 51
PCBs 209
FCCDB 4268
INTeraction).
2.8. Molecular docking with Autodock Vina software
Molecular docking experiments were performed with Autodock Vina
1.1.2 using default settings (Trott and Olson, 2009). The search space
was included in a box of 24 × 24 × 24 Å, centred on the binding site of
the ligands as mentioned before. The side chain flexibility was allowed
for the same residues defined in the GOLD docking. The ligand amide
and backbone flexibility were allowed.
3. Results and discussion
The foodchem DB has been also designed to accelerate computa­
tional applications since it stores not only regulative information but
also chemical-physical properties and three-dimensional structures.
Very careful attention has been made to ensure the correctness of the 3D
structure to the CAS RN. Thus, it has been conceived for a different
purpose compared to the FPF database which does not contain all the
chemical-physical information used in the foodchem DB and it does not
store the three-dimensional structure. Moreover, our database has been
written in SQL and NoSQL language with the purpose to make it avail­
able to the scientific community through a website interface where the
user can make searches and extract information. Using our database, the
three-dimensional structures of 8091 substances, belonging to different
sub-classes (Table 2), has been extracted and all these molecules have
been screened using a molecular docking approach in order to identify
the compounds having the capability to bind the thirty-one nuclear re­
ceptors. This method allows to screen the substances which have the
most probable physical-chemical characteristics to act as endocrine
disruptors.
Two different docking software and four different scoring functions
have been used as in our previous papers (Francesca Cavaliere et al.,
2020) (Spaggiari et al., 2021). Thus, for each receptor and for each food
contact chemical, four values have been obtained. In humans, there are
48 nuclear receptors, but many of these remain “orphans” as their
endogenous ligands are yet to be determined. For this reason, if the
endogenous ligand is known, the relative binding affinity (RBA) of each
molecule was calculated using it as a reference compound. On the other
hand, all the endogenous and no-endogenous co-crystallized ligands
were docked against the respective nuclear receptors to obtain a refer­
ence value. A cut-off value was selected for each four docking values: i) a
cut-off of 50 for GoldScore; ii) a cut-off of 30 for ChemScore; iii) a cut-off
of − 7 for Autodock (affinity); and iv) a cut-off of 500 for HintScore.
To reach a consensus scoring prediction, a robust statistical method
has been used and it is explained in more detail below.
As training dataset, the crystallographic structures available from
P. Cozzini et al. Chemosphere 292 (2022) 133422

Fig. 1. Results obtained from the robust multivariate statistical procedure. The 31 NRs are on the x-axis, while the number of the molecules (%) is on the ordinate.
The molecules with a score smaller than 0.3 are highlighted in green (A), the molecules with a score between 0.3 and 0.8 are highlighted in yellow (B), the molecules
with a score greater than 0.8 are highlighted in red (C), while the outliers are highlighted in grey (D). (For interpretation of the references to colour in this figure
legend, the reader is referred to the Web version of this article.)

PDB of all ligand-NR complexes were considered. All ligands bound to ∑4

the corresponding receptor were extracted and docked into the ligand-
binding pocket to obtain the corresponding four scoring values. As for
the food contact chemical data, every single value was used to calculate
the relative binding activity considering the natural ligand as a reference
compound:
food contact chemical score
Relative Binding Affinity(RBA)n =
reference compound score
where n is the number of scoring functions.
However, since the distribution data is non-normal for the potential
presence of some outliers, a robust multivariate method was used to
detect atypical values. In fact, it is well-known that the presence of
atypical values can affect the results of any statistical analysis especially
when the number of observations is large. Using a confidence level of
simultaneous 1%, we removed only values that were very far from the
general bulk of the data. After the outlier removal, the values were
rescaled in the domain [0 1] setting a score equal to 1 when it was larger
than the value of the natural ligand. The degree of dispersion of the four
rescaled values (X1 , …,X4 ) has been considered by normalizing them in
order to obtain four new variables (Z1 , …,Z4 ) with 0 mean and variance
equal to 1. After that, a principal component analysis was used on the
four new variables to identify a weight coefficient for each scoring
function (w1, w2, w3, w4) in such a way that the explained variance of
4
∑
the original variable is as large as possible (wj ≥ 0 and w2j = 1). We
j=1
obtained a weight value of 0.12–0.94 – 0.14–0.29, for GoldScore,
HintScore, ChemScore, and Autodock (affinity), respectively.
As for the training dataset, the relative binding affinity of each
molecule and scoring function has been rescaled in the [0 1] domain
after the outlier removal. To consider the different degrees of dispersion
of the new rescaled variables, we standardized them to obtain four new
variables. Since the purpose of the analysis was to combine the four
scores into a single consensus score prediction, the final scores for the i-
th food contact chemical have been obtained as:
j=1 xij wj
∑4 i = 1, 2, …, n
j=1 wj
where n is the total number of food contact chemicals.
At the end, the results have been divided into three cases based on
their score: i) the molecules with a score between 0.0 and 0.3 are
considered weak ligands since they interact with the corresponding
nuclear receptor with a binding affinity that is 70% (or more) lower than
the natural ligand (Fig. 1A); ii) the molecules with a score between 0.3
and 0.8 are considered medium interactor compared to the natural
ligand (Fig. 1B); iii) the molecules with a score between 0.8 and 1.0 are
judged as high interactor since they are able to bind the corresponding
nuclear receptor with a binding affinity that is more than 80% of the
natural ligand (Fig. 1C). This latter case also includes the molecules that
can interact with the nuclear receptor with a binding affinity greater
than the natural ligand. Thus, all food contact chemicals falling in this
class may be considered as substances of very high concern and should
be the first compounds to analyse with further experimental methods in
order to re-evaluate their use in the food industry.
Fig. 1 shows the percentage of molecules that can interfere with the
endocrine system receptors highlighting their abundance in each spe­
cific nuclear receptor. If we focus on the single nuclear receptor, we can
underline that more than 50% of food contact chemicals are good
interactors of liver X receptor β (LXRβ), pregnane X receptor (PXR),
progesterone receptor (PR), farnesoid X receptor (FXR), retinoic acid-
related orphan receptor γ (RORγ), and peroxisome proliferator-
activated receptor α (PPARα). In fact, LXRβ is the nuclear receptor
with the highest number of food contact chemicals that fall in the high
interactor group, and thus, it is likely the receptor most affected by the
presence of these compounds in our body.
Considering Fig. 1D, we found almost the same number of outlier
molecules in each nuclear receptor except for the estrogen-related re­
ceptor α (ERRα). This is not surprising since outlier molecules were
generally substances having a high volume compared to the ligand-
binding pocket of nuclear receptors. In fact, due to atom-atom clashes,

5
P. Cozzini et al.
Fig. 2. The percentage of molecules able to bind more than 15 nuclear receptors with high (≥0.8), medium (0.3–0.8), and low binding affinity (<0.3) considering
each class of food contact chemicals.
the molecular docking scores were far away from the normal trend.
Thus, considering that the volume of the ligand-binding pocket of ERRα
is only about 80 Å3 (against the ~300 Å3 of the most nuclear receptor,
excluding the PPAR family), it may be plausible to find a higher number
of outliers.
As the second step of our analysis, we turned our attention on which
class of food contact chemicals have the greater number of molecules
able to interfere with the endocrine system. Thus, we counted the
number of molecules belonging to each class that can interact with more
than 50 percent of nuclear receptors with high, medium, and low
binding affinity. As we can see in Fig. 2, almost the totality of dioxins,
furans, and PCBs molecules can interact with more than 15 nuclear re­
ceptors with high binding affinity, following by the pesticides and
phthalates sub-classes.
The impact of this finding highlights the potential capability of these
molecules to cause a very broad endocrine effect on the human body.
Considering the medium interactors, a great number of flavourings,
bisphenols, and FCCDBs fall in this group. The single compound in the
acrylamide class is also able to interact with more than fifteen nuclear
receptors with medium binding affinity. On the other site, food additives
and mycotoxins are more selective in their interaction with nuclear re­
ceptors, and just a few numbers of molecules can interact with high
affinity to more than 50 percent of NRs.
4. Conclusion
One of the reasons that undermine in silico approaches is the avail­
ability of highly curated databases from which to retrieve and download
the three-dimensional structure. This is most relevant in the food context
due to the presence of salt and mixture components. In fact, it is frequent
on the web to find mixture or salt substances associated with the CAS RN
of the main compound. In the present work, we created a database with
a high level of data curation from which to retrieve chemical, structure,
and regulative information about all food contact chemicals.
Using our foodchem database, we screened 8091 food contact
chemicals against 31 nuclear receptors with the aim to identify the
molecules that require major attention about their safety for the human
body. In the food context, wet experiments are the most used and
accepted methods and, thus, there is often a mistrust about the reli­
ability of computational techniques. However, dry experiments also
have their drawbacks. For example, the compound 4′ -Methox­
yacetophenone (CAS RN: 100-06-1), which is used as an additive and
6
Chemosphere 292 (2022) 133422
flavouring compound, and it is also included in the Food Contact
Chemical DB (FCCDB), has two different predicted activities for its
capability to act as an agonist for the estrogen receptor α. In fact, in the
Tox21 project (Richard et al., 2021), the quantitative high-throughput
screening assay (qHTS) identifies 4′ -Methoxyacetophenone both as
active and inactive for its agonist activity on ERα. In light of this, we
think that there is not an approach that can be judged as better than
another, but all are equally valid and should be considered together.
Thus, the present work should not be seen as an opposing method to
classical in vitro and in vivo tests, but it should be considered as a useful
and preliminary method to screen a huge number of molecules in a cost
and time-effective manner. In fact, using our robust computational
method, we screened a large volume of molecules against the nuclear
receptor family in a relatively short time when compared to the time
needed for in vitro and in vivo experiments.
Author contribution statement
Pietro Cozzini – Conceptualization, Methodology, Project adminis­
tration, Resources, Supervision, Writing reviewing/editing, Giulia
Spaggiari – Data curation, Formal analysis, Investigation, Methodology,
Validation, Writing – original draft, Francesca Cavaliere – Data curation,
Formal analysis, Investigation, Methodology, Software Development,
Validation, Writing – original draft.Marco Riani & Gianluca Morelli –
Statistical methods and software development
Declaration of competing interest
The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to influence
the work reported in this paper.
References
Bissantz, C., Folkers, G., Rognan, D., 2000. Protein-based virtual screening of chemical
databases . 1 . Evaluation of different docking/scoring combinations. J. Med. Chem.
43, 4759–4767. https://doi.org/10.1021/jm001044l.
Cavaliere, F., Lorenzetti, S., Cozzini, P., 2020. Molecular modelling methods in food
safety: bisphenols as case study. Food Chem. Toxicol. 137 https://doi.org/10.1016/j.
fct.2020.111116.
Cavaliere, Francesca, Spaggiari, G., Cozzini, P., 2020. Molecular Docking for Computer-
Aided Drug Design: Fundamentals, Techniques, Resources and Applications.
Dall’Asta, C., 2016. Mycotoxins and nuclear receptors: a still underexplored issue. Nucl.
Recept. Res. 3 https://doi.org/10.11131/2016/101204.
P. Cozzini et al.
De Coster, S., Van Larebeke, N., 2012. Endocrine-disrupting chemicals: associated
disorders and mechanisms of action. J. Environ. Public Health 2012. https://doi.
org/10.1155/2012/713696.
Desvergne, B., Feige, J.N., Casals-Casas, C., 2009. PPAR-mediated activity of phthalates:
a link to the obesity epidemic? Mol. Cell. Endocrinol. 304, 43–48. https://doi.org/
10.1016/j.mce.2009.02.017.
Fucic, A., Gamulin, M., Ferencic, Z., Katic, J., Krayer Von Krauss, M., Bartonova, A.,
Merlo, D.F., 2012. Environmental exposure to xenoestrogens and oestrogen related
cancers: reproductive system, breast, lung, kidney, pancreas, and brain. Environ.
Heal. A Glob. Access Sci. Source 11, 1–9. https://doi.org/10.1186/1476-069X-11-
S1-S8.
Gore, A.C., Chappell, V.A., Fenton, S.E., Flaws, J.A., Nadal, A., Prins, G.S., Toppari, J.,
Zoeller, R.T., 2015. EDC-2: the endocrine society’s second scientific statement on
endocrine-disrupting chemicals. Endocr. Rev. 36, E1–E150. https://doi.org/
10.1210/er.2015-1010.
Grulke, C.M., Williams, A.J., Thillanadarajah, I., Richard, A.M., 2019. EPA’s DSSTox
database: history of development of a curated chemistry resource supporting
computational toxicology research. Comput. Toxicol. 12, 100096 https://doi.org/
10.1016/j.comtox.2019.100096.
Hersey, A., Chambers, J., Bellis, L., Patrícia Bento, A., Gaulton, A., Overington, J.P.,
2015. Chemical databases: curation or integration by user-defined equivalence?
Drug Discov. Today Technol 14, 17–24. https://doi.org/10.1016/j.
ddtec.2015.01.005.
Kabir, E.R., Rahman, M.S., Rahman, I., 2015. A review on endocrine disruptors and their
possible impacts on human health. Environ. Toxicol. Pharmacol. 40, 241–258.
https://doi.org/10.1016/j.etap.2015.06.009.
Kavlock, R.J., Daston, G.P., DeRosa, C., Fenner-Crisp, P., Gray, L.E., Kaattari, S.,
Lucier, G., Luster, M., Mac, M.J., Maczka, C., Miller, R., Moore, J., Rolland, R.,
Scott, G., Sheehan, D.M., Sinks, T., Tilson, H.A., 1996. Research needs for the risk
assessment of health and environmental effects of endocrine disrupters: a report of
the U.S. EPA-sponsored workshop. Environ. Health Perspect. 104, 715–740. https://
doi.org/10.1289/ehp.96104s4715.
Luccio-Camelo, D.C., Prins, G.S., 2011. Disruption of androgen receptor signaling in
males by environmental chemicals. J. Steroid Biochem. Mol. Biol. 127, 74–82.
https://doi.org/10.1016/J.JSBMB.2011.04.004.
Chemosphere 292 (2022) 133422
Morris, G.M., Lim-Wilby, M., 2008. Molecular docking. Methods Mol. Biol. 443,
365–382. https://doi.org/10.1007/978-1-59745-177-2_19.
Odermatt, A., Gumy, C., 2008. Glucocorticoid and mineralocorticoid action: why should
we consider influences by environmental chemicals? Biochem. Pharmacol. 76,
1184–1193. https://doi.org/10.1016/j.bcp.2008.07.019.
Petrakis, D., Vassilopoulou, L., Mamoulakis, C., Psycharakis, C., Anifantaki, A.,
Sifakis, S., Docea, A.O., Tsiaoussis, J., Makrigiannakis, A., Tsatsakis, A.M., 2017.
Endocrine disruptors leading to obesity and related diseases. Int. J. Environ. Res.
Public Health 14, 1–18. https://doi.org/10.3390/ijerph14101282.
Richard, A.M., Huang, R., Waidyanatha, S., Shinn, P., Collins, B.J., Thillainadarajah, I.,
Grulke, C.M., Williams, A.J., Lougee, R.R., Judson, R.S., Houck, K.A., Shobair, M.,
Yang, C., Rathman, J.F., Yasgar, A., Fitzpatrick, S.C., Simeonov, A., Thomas, R.S.,
Crofton, K.M., Paules, R.S., Bucher, J.R., Austin, C.P., Kavlock, R.J., Tice, R.R., 2021.
The Tox21 10K compound library: collaborative chemistry advancing toxicology.
Chem. Res. Toxicol. 34, 189–216. https://doi.org/10.1021/acs.
chemrestox.0c00264.
Safe, S., 2004. Endocrine disruptors and human health: is there a problem. Toxicology
205, 3–10. https://doi.org/10.1016/j.tox.2004.06.032.
Schug, T.T., Janesick, A., Blumberg, B., Heindel, J.J., 2011. Endocrine disrupting
chemicals and disease susceptibility. J. Steroid Biochem. Mol. Biol. 127, 204–215.
https://doi.org/10.1016/j.jsbmb.2011.08.007.
Spaggiari, G., Iovine, N., Cozzini, P., 2021. In silico prediction of the mechanism of
action of pyriproxyfen and 4′ -oh-pyriproxyfen against a. Mellifera and h. sapiens
receptors. Int. J. Mol. Sci. 22 https://doi.org/10.3390/ijms22147751.
Teramoto, R., Fukunishi, H., 2007. Supervised consensus scoring for docking and virtual
screening. J. Chem. Inf. Model. 47, 526–534. https://doi.org/10.1021/ci6004993.
Trott, O., Olson, A.J., 2009. Software news and update AutoDock Vina: improving the
speed and accuracy of docking with a new scoring function, efficient optimization,
and multithreading. J. Comput. Chem. 31, 455–461.
Wang, R., Lu, Y., Wang, S., 2003. Comparative evaluation of 11 scoring functions for
molecular docking. J. Med. Chem. 46, 2287–2303. https://doi.org/10.1021/
jm0203783.